Drones play an important role in the development of remote sensing and intelligent surveillance. Due to limited onboard computational resources, drone-based object detection still faces challenges in actual applications. By studying the balance between detection accuracy and computational cost, we propose a novel object detection algorithm for drone cruising in large-scale maritime scenarios. Transformer is introduced to enhance the feature extraction part and is beneficial to small or occluded object detection. Meanwhile, the computational cost of the algorithm is reduced by replacing the convolution operations with simpler linear transformations. To illustrate the performance of the algorithm, a specialized dataset composed of thousands of images collected by drones in maritime scenarios is given, and quantitative and comparative experiments are conducted. By comparison with other derivatives, the detection precision of the algorithm is increased by 1.4%, the recall is increased by 2.6% and the average precision is increased by 1.9%, while the parameters and floating-point operations are reduced by 11.6% and 7.3%, respectively. These improvements are thought to contribute to the application of drones in maritime and other remote sensing fields.

Robust detection and recognition of water surfaces are critical for autonomous navigation of unmanned surface vehicles (USVs), since any none-water region is likely an obstacle posing a potential danger to the sailing vehicle. A novel water region visual detection method is proposed in this paper. First, the input image pixels are clustered into different regions and each pixel is assigned a label tag and a confidence value by adaptive multistage segmentation algorithm. Then the resulting label map and associated confidence map are fed into a convolutional neural network (CNN) as training samples to train the network online. Finally, the online trained CNN is used to segment the input image again but with greater precision and stronger robustness. Compared with other deep-learning image segmentation algorithms, the proposed method has two advantages. Firstly, it dispenses with the need of manual labeling training samples which is a costly and painful task. Secondly, it allows real-time online training for CNN, making the network adaptive to the navigational environment. Another contribution of this work relates to the training process of neuro network. An effective network training method is designed to learn from the imperfect training data. We present the experiments in the lake with a various scene and demonstrate that our proposed method could be applied to recognize the water region in the unknown navigation environment automatically.

The prototypes of autonomous vessels are expected to come into service within the coming years, but safety concerns remain due to complex traffic and natural conditions (e.g., Yangtze River). However, the response of autonomous vessels to potential accidents is still uncertain. The accident prevention for autonomous vessels is unconvincing due to the lack of objective studies on the causation analysis for maritime accidents. This paper constitutes an attempt to cover the aforementioned gap by studying the potential causations for maritime accidents in the Yangtze River by using a Bayesian-based network training approach. More than two hundred accidents reported between 2013 and 2019 in the Yangtze River are collected. As a result, a Bayesian network (BN) is successfully established to describe the causations among different risk influencing factors. By analysing the BN, this study reveals that the occurrence of maritime accidents (e.g., collision, grounding) can be expected to reduce with the development of autonomous vessels as the crews are removed. However, the extent of the consequences from some accidents (e.g., fire, critical weathers) could be more serious than conventional ones. Therefore, more attention and thoughts are needed to ensure the safe navigation of autonomous vessels in the Yangtze River.